SecureCode: Security-Aware AI Training Dataset

Arxiv pdf 2025-12-01T00:00:00
arXiv Paper — PDF not available. Only the Executive Summary is available here. To read or download the full paper, visit the arXiv abstract page.

Abstract

AI coding assistants produce vulnerable code in 45% of security-relevant scenarios [Veracode, 2025], yet no public training dataset teaches both traditional web security and AI/ML-specific defenses in a format suitable for instruction tuning. We present SecureCode, a production-grade dataset of 2,372 multi-turn security training examples spanning two domains: web application security (1,625 examples covering the OWASP Top 10 2021 across 11 languages and 9 frameworks) and AI/ML security (747 examples covering all 10 OWASP LLM Top 10 2025 categories across 40+ frameworks including LangChain, OpenAI, and HuggingFace). A 2026 grounding audit (Section 5.6) independently re-verified every reference and corrected the datasets grounding claims: we report the fraction of examples tied to a documented incident or verified CVE rather than asserting complete grounding. Every example follows a 4-turn conversational structure—feature request, vulnerable and secure implementations with attack demonstrations, advanced probing, and defense-in-depth operational guidance—designed for direct use in instruction tuning pipelines.

Loading executive summary...

LINK COPIED TO CLIPBOARD